Add SwiftFormer, SHViT, StarNet, FasterNet and GhostNetV3

[brianhou0208]

New Model

• (ICCV2023) SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications
• (CVPR2023) Run, Don’t Walk: Chasing Higher FLOPS for Faster Neural Networks
• (CVPR2024) SHViT: Single-Head Vision Transformer with Memory Efficient
• (CVPR2024) Rewrite the Stars
• GhostNetV3: Exploring the Training Strategies for Compact Models

Model Request

resolve #2450

Result

Model	Acc@1	MACs	Params	GPU Throughput	CPU Throughput	iOS NPU Latency	Android NPU Latency
swiftformer_xs	75.586	603.77M	3.48M	3356.03	146.25	0.63	3.288
swiftformer_s	78.466	986.19M	6.09M	2519.29	123.62	0.76	3.386
swiftformer_l1	80.902	1.60G	12.06M	1832.04	100.77	1.07	3.810
swiftformer_l3	83.000	4.01G	28.49M	969.61	43.39	1.88	4.700
fasternet_t0	71.904	335.69M	3.91M	8924.14	461.35	0.50	0.256
fasternet_t1	76.210	850.94M	7.60M	4939.89	262.55	0.67	0.384
fasternet_t2	78.896	1.90G	14.98M	2548.34	133.69	1.04	0.641
fasternet_s	81.290	4.55G	31.18M	1219.00	62.62	1.88	1.257
fasternet_m	82.966	8.72G	53.52M	605.40	28.72	-	2.456
fasternet_l	83.506	15.50G	93.47M	381.80	17.43	-	4.505
shvit_s1	72.778	241.47M	6.31M	13566.14	494.77	0.60	0.461
shvit_s2	75.192	365.58M	11.45M	10214.81	413.26	0.81	0.523
shvit_s3	77.340	600.83M	14.21M	6676.91	299.91	0.96	0.625
shvit_s4	79.378	986.63M	16.55M	4149.87	177.38	1.09	0.798
starnet_s1	73.564	422.89M	2.87M	3859.64	233.59	0.48	0.471
starnet_s2	74.660	545.05M	3.68M	3888.33	203.43	0.51	0.438
starnet_s3	77.380	755.38M	5.75M	2760.39	150.84	0.66	0.548
starnet_s4	78.838	1.05G	7.48M	1938.80	112.58	0.81	0.720
ghostnetv3_100	76.928	168.00M	6.15M	3467.852	168.138	-	0.668

• Only SHViT S4 uses (256, 256) input image size
• SwiftFormer weight is on google drive

Param / MACs / Throughput

NPU Latency

• iOS NPU: Apple iPhone 14 Pro Max / iOS 18.5 / A16
• Android NPU: Samsung S24 / Android 14 / Snapdragon® 8 Gen 3 | SM8650

• iOS latency reported for iPhone 14 Pro Max (iOS 18.5) uses the benchmark tool from Xcode 16.3

• Android latency reported for Samsung Galaxy S24 (Android 14) uses the benchmark tool from Qualcomm® AI Hub Models

Measure Android Latency

import torch
import timm
from timm.utils.model import reparameterize_model
import qai_hub as hub


def latency_android(name, img_size=224):
    torch_model = timm.create_model(name)
    torch_model.eval()
    torch_model = reparameterize_model(torch_model)
    input_shape = (1, 3, img_size, img_size)
    example_input = torch.rand(input_shape)
    traced_model = torch.jit.trace(torch_model, example_input)
    compile_job = hub.submit_compile_job(
        model=traced_model,
        device=hub.Device("Samsung Galaxy S24 (Family)"),
        input_specs=dict(image=input_shape),
    )
    target_model = compile_job.get_target_model()
    profile_job = hub.submit_profile_job(
        model=target_model,
        device=hub.Device("Samsung Galaxy S24 (Family)"),
    )
    profile_result = profile_job.download_profile()
    latency_ms = profile_result["execution_summary"]["estimated_inference_time"] / 1000
    print(f"{name} 📊 Latency: {latency_ms:.3f} ms")


if __name__ == '__main__':
    name_list = timm.list_models('shvit*')
    print(name_list)
    for name in name_list:
        latency_android(name)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.